Probabilistic Open Set Recognition

Abstract

Real-world tasks in computer vision, pattern recognition and machine learning often touch upon the open set recognition problem: multi-class recognition with incomplete knowledge of the world and many unknown inputs. An obvious way to approach such problems is to develop a recognition system that thresholds probabilities to reject unknown classes. Traditional rejection techniques are not about the unknown; they are about the uncertain boundary and rejection around that boundary. Thus traditional techniques only represent the “known unknowns”. However, a proper open set recognition algorithm is needed to reduce the risk from the “unknown unknowns”. This dissertation examines this concept and finds existing probabilistic multi-class recognition approaches are ineffective for true open set recognition. We hypothesize the cause is due to weak adhoc assumptions combined with closed-world assumptions made by existing calibration techniques. Intuitively, if we could accurately model just the positive data for any known class without overfitting, we could reject the large set of unknown classes even under this assumption of incomplete class knowledge. For this, we formulate the problem as one of modeling positive training data by invoking statistical extreme value theory (EVT) near the decision boundary of positive data with respect to negative data. We provide a new algorithm called the $P_{I}$-SVM for estimating the unnormalized posterior probability of class inclusion. This dissertation also introduces a new open set recognition model called Compact Abating Probability (CAP), where the probability of class membership decreases in value (abates) as points move from known data toward open space. We show that CAP models improve open set recognition for multiple algorithms. Leveraging the CAP formulation, we go on to describe the novel Weibull-calibrated SVM (W-SVM) algorithm, which combines the useful properties of statistical EVT for score calibration with one-class and binary support vector machines. Building from the success of statistical EVT based recognition methods such as $P_{I}$-SVM and W-SVM on the open set problem, we present a new general supervised learning algorithm for multi-class classification and multi-class open set recognition called the Extreme Value Local Basis (EVLB). The design of this algorithm is motivated by the observation that extrema from known negative class distributions are the closest negative points to any positive sample during training, and thus should be used to define the parameters of a probabilistic decision model. In the EVLB, the kernel distribution for each positive training sample is estimated via an EVT distribution fit over the distances to the separating hyperplane between positive training sample and closest negative samples, with a subset of the overall positive training data retained to form a probabilistic decision boundary. Using this subset as a frame of reference, the probability of a sample at test time decreases as it moves away from the positive class. Possessing this property, the EVLB is well-suited to open set recognition problems where samples from unknown or novel classes are encountered at test. Our experimental evaluation shows that the EVLB provides a substantial improvement in scalability compared to standard radial basis function kernel machines, as well as $P_{I}$-SVM and W-SVM, with improved accuracy in many cases. We evaluate our algorithm on open set variations of the standard visual learning benchmarks, as well as with an open subset of classes from Caltech 256 and ImageNet. Our experiments show that $P_{I}$-SVM, WSVM and EVLB provide significant advances over the previous state-of-the-art solutions for the same tasks.